A comparison of gradient-based algorithms for echo compensation with decorrelating properties

Cancelling echoes by using the normalized least mean square (NLMS) algorithm has been state of the art for many years. In acoustical echo compensation, however, it is common to estimate more than 1000 parameters resulting in a too slow convergence when driven by speech signals. In order to overcome this drawback, a lot of modifications have been published in the last years, all having one goal: to decorrelate the driving process. Beginning with a deterministic approach we show that all these different ideas can be arranged in one scheme, allowing a uniform normalization. The different properties of the several algorithms are then obvious. A comparison of some algorithms with 2N-4N complexity is presented. Surprisingly, all algorithms do not work perfectly for a large compensator filter length and speech as input process